STEVE EISMAN: AI revenue across the supply chain is concentrated in OpenAI and Anthropic, so the whole buildout depends on two loss-making companies

The Gist

Steve Eisman says most AI money in the chip and cloud world traces back to OpenAI and Anthropic. If those two stumble, everyone who sells to them does too. This is a steelman: Steve Eisman's arguments are reconstructed in their strongest fair form, not endorsed.

Conclusion

Because AI revenue at Nvidia, the hyperscalers and Oracle is concentrated in OpenAI and Anthropic, the AI buildout currently depends on the health of two companies that do not yet reliably generate cash, and that concentration, more than any single neocloud, is the system's main vulnerability.

Premises

1. Five direct customers accounted for 70 percent of Nvidia's accounts receivable in its latest quarterly filing.
2. Analyst estimates put OpenAI and Anthropic at roughly 70 percent of the hyperscalers' AI revenue and roughly a quarter or more of some cloud businesses.
3. About half of Oracle's contracted backlog is tied to OpenAI.
4. OpenAI and Anthropic made large commitments with many suppliers, and the buildout is sized to those commitments.
5. A business whose revenue depends on one or two customers is exposed to the loss of either.

Assumptions

• Nvidia's direct customers are mostly cloud providers with many end customers, and the share-of-revenue figures for the hyperscalers are analyst estimates the companies have not confirmed.
• Anthropic reported positive cash flow in the third quarter of 2026, which qualifies "bleeds cash" for that company.
• Eisman expects the concentration to matter less in three to four years as other buyers of compute emerge.

Analysis

Argument type: Inductive.

Potential Fallacies

• Unsupported comparative claim (Conclusion (final clause: 'more than any single neocloud, is the system's main vulnerability')) — The conclusion ranks lab concentration as the system's main vulnerability, above any single neocloud. No premise compares it with neocloud leverage, power constraints, circular financing, chip supply or demand durability. The ranking is asserted rather than derived. Neocloud risk and lab-concentration risk are also probably correlated, since neoclouds often depend on the same labs.
• Proxy evidence / conflating direct customers with end demand (Premise 1 to the conclusion) — Five direct customers making up 70 percent of Nvidia's receivables shows buyer concentration at the hardware layer. By A1 those buyers are mostly clouds with many end customers, and they also buy for internal workloads. The figure therefore does not show that Nvidia depends on OpenAI and Anthropic. Receivables also reflect billing timing rather than revenue share. The inference to lab dependence has to pass through P2, which rests on estimates.
• Exposure-versus-vulnerability conflation (missing failure probability) (Premise 5 to the conclusion) — P5 is close to a truism: any concentrated business is exposed to its customers. Exposure becomes vulnerability only when the probability of a customer failing or cutting back is meaningful and replacement demand is limited. The argument gives no estimate of either lab's failure risk, contract enforceability, funding runway or the resale value of freed capacity. It also treats risk as a binary 'loss' of a customer, when renegotiation or resizing is more likely.
• Unsupported conclusion element / inconsistent characterization (Title and conclusion versus A2) — The claim that the two firms 'do not yet reliably generate cash' has no supporting premise, and the headline's 'loss-making' framing goes beyond A2, which says Anthropic reported positive cash flow. The conclusion's 'not yet reliably' is more defensible than the headline, since one quarter is thin evidence either way. Accounting losses, cash burn and insolvency risk are also distinct, and the argument treats them as interchangeable. Lumping the two labs together hides differences in their…
• Generalizing from the most exposed node (Premises 2 and 3 to 'the whole buildout') — Oracle, with about half of its backlog tied to OpenAI, is the most exposed participant. Its profile is used to characterize 'Nvidia, the hyperscalers and Oracle' together. Hyperscalers have large diversified businesses and internal AI demand, so average exposure across the chain is likely lower than Oracle's. Percentages of AI revenue are also not percentages of total revenue or of total demand for the buildout.
• Reliance on unconfirmed estimates presented as settled (Premise 2) — The central hyperscaler figure is an analyst estimate with undisclosed methodology and an undefined 'AI revenue' denominator. Estimates from a small pool of sources may share assumptions and so are not independent. A 'confirmed' label in the brief means the estimate was reported, not that the underlying quantity is verified. A1 concedes this, but the conclusion's confidence still outruns the status of the evidence.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — Concentration is not fragility. The labs have raised very large amounts of capital, have strategic backers (several of them the suppliers themselves), and are growing quickly. Hyperscalers are cash-rich and diversified, with internal AI demand that would continue even if a lab stumbled. A lab shortfall would more likely mean renegotiation, resizing or repricing of capacity than a collapse.
• Conclusion (High impact) — The more serious fragility may lie elsewhere: debt-financed neoclouds and off-balance-sheet vehicles, circular financing in which suppliers fund the labs that buy from them, power and supply constraints, or weakening model-pricing economics. The argument compares concentration only with a single neocloud, not with these systemic alternatives.
• Premise 1 (High impact) — Receivables concentration is routine for a supplier selling to a few hyperscalers and OEMs/ODMs, and hyperscalers buy GPUs for search, ads, recommendation and their own products. The 70 percent figure therefore says little about OpenAI or Anthropic.
• Conclusion (Medium impact) — Anthropic's reported positive cash flow (A2), together with OpenAI's deep backing, shows the two labs are in different financial situations. The 'two loss-making companies' picture is partly outdated. Only one lab faltering would show resilience rather than systemic fragility.
• Premise 4 (Medium impact) — GPU and data-center capacity is fungible in a capacity-constrained market. If one lab cut commitments, others would likely absorb the compute at some price, as A3 itself implies, so stranded capacity is less likely than the argument suggests.
• Premise 3 (Medium impact) — Oracle's backlog is a legal obligation. Take-or-pay terms, prepayments or parent and investor support could substantially reduce default risk relative to the headline share.
• Conclusion (Medium impact) — Taken at face value, the reasoning would mark every industry with concentrated buyers (suppliers to Apple, Boeing or defense primes) and every venture-funded buildout as systemically fragile. Concentration is a vulnerability only when failure is probable and substitutes are absent.

Suggested Improvements

• Probability and severity of failure — Add premises on the labs' funding runway, fundraising access and revenue growth, and estimate the likelihood and size of a pullback. Model supplier revenue under a 25 to 50 percent cut in lab spending. Exposure alone (P5) does not make a vulnerability. Risk requires both probability and impact, and the argument supplies only impact.
• Comparative ranking — Either drop 'main vulnerability' for 'a significant, time-limited vulnerability', or support it by comparing lab concentration with neocloud leverage, circular financing, power constraints and demand durability on a stated metric. The comparative claim currently has no supporting premise, and the neocloud and lab risks likely overlap.
• Bridging Nvidia to the labs — Look through Nvidia's direct customers to end demand, or present P1 explicitly as context about buyer concentration rather than as evidence of lab dependence. Separate the receivables figure from the lab-share estimates. This removes the apparent conflation of two different '70 percent' statistics and keeps the strongest data points from being discounted.
• Contract and financing structure — Add evidence on take-or-pay terms, termination rights, prepayments, parent guarantees, and vendor and equity financing links between suppliers and labs. These determine whether stress is transmitted, absorbed or amplified, and they cover the circularity issue that could make the risk worse than stated.
• Differentiate the two labs — Treat OpenAI and Anthropic separately on cash position and funding access, and assess how correlated their failures would be. Align the headline with A2 and the conclusion's 'not yet reliably' wording. They have different financial profiles, and pooling them overstates the risk. The headline currently goes beyond the argument's own assumption.
• Evidence for P4 and non-lab demand — Quantify what share of capex and committed capacity is attributable to the two labs, versus internal hyperscaler, enterprise and sovereign demand, and how quickly the latter is growing. 'Sized to those commitments' is the key bridge to whole-buildout dependence and is currently unsupported.
• Time scope — State explicitly that the claim concerns the next three to four years, and name indicators or triggers (funding closes, contract deferrals, backlog conversion, non-lab demand growth) that would confirm or disconfirm it. A3 makes the risk transitional. Without defined indicators the thesis is hard to falsify, since it can be vindicated either by a crisis or by the concentration fading.

Scenario Tests

• OpenAI cannot raise capital on the scale needed and defers or cancels a large share of its Oracle and cloud commitments. (Supports) — Oracle, with half its backlog tied to OpenAI, and the leveraged neoclouds would be hit hardest. This confirms the concentration thesis at the most exposed nodes. Hyperscalers and Nvidia would be affected less directly, which shows the dependence is uneven across the chain.
• Enterprise, sovereign, consumer and other-lab demand for compute grows faster than expected, diluting lab share well before three to four years. (Challenges) — The concentration risk fades early and is revealed as a transitional feature of an early market. This is consistent with A3 and undermines 'main vulnerability'.
• Only one lab (for example OpenAI) falters while Anthropic stays cash-flow positive and absorbs some of the freed capacity. (Challenges) — The 'two companies' framing breaks and the system shows partial resilience. The vulnerability lies in one lab's financing, not in the pair.
• A leveraged neocloud defaults or credit markets seize while both labs keep paying. (Challenges) — Financing structure rather than customer identity is the binding constraint, contradicting the ranking of concentration above neocloud risk. In practice the two risks overlap, since neocloud debt is backed by lab contracts.
• A lab renegotiates, stretches or resizes its commitments without failing, or capacity is resold to other buyers at lower prices. (Neutral) — Supplier revenue and valuations would take a hit and the exposure is real, but the outcome is a repricing, not a systemic break. This is the most likely stress path and the one the argument does not model.
• Investors and lenders reprice lab exposure after the concentration narrative spreads, raising the labs' funding costs. (Supports) — Reflexivity could make the vulnerability partly self-fulfilling, but it also shows the argument is a static snapshot of a dynamic, feedback-driven system.

Coherence & Relevance

The argument coheres well as a case that AI compute demand is materially concentrated in two frontier labs, and the stated assumptions (A1 to A3) make that case honest about its limits. It coheres less well at the conclusion's further steps. No premise establishes that both firms are cash-fragile (A2 cuts against it), and none compares lab concentration with other vulnerabilities. The step from concentration to a 'main' systemic vulnerability also needs evidence about failure probability and contagion channels that is not provided. The conclusion is best read as 'a significant but time-limited concentration risk' rather than 'the main vulnerability'.

• Five direct customers accounted for 70 percent of Nvidia's accounts receivable in its latest quarterly filing. (Weak) — Describes direct-customer and billing concentration, not dependence on OpenAI or Anthropic. By A1 the customers are mostly diversified clouds. The link to the conclusion runs only through P2, which is an estimate.
• Analyst estimates put OpenAI and Anthropic at roughly 70 percent of the hyperscalers' AI revenue and roughly a quarter or more of some cloud businesses. (Strong) — Directly addresses lab concentration at the hyperscalers, but the denominator ('AI revenue') is undefined and the estimates are unconfirmed. It does not translate AI-revenue share into dependence of the whole buildout.
• About half of Oracle's contracted backlog is tied to OpenAI. (Strong) — Strongly supports Oracle-specific exposure. It does not generalize to Nvidia or the hyperscalers, and backlog quality and collectability are not addressed.
• OpenAI and Anthropic made large commitments with many suppliers, and the buildout is sized to those commitments. (Strong) — The right bridge between customer concentration and buildout dependence, but unsourced and unquantified. It does not address binding versus non-binding commitments, other demand sources, or capacity substitutability.
• A business whose revenue depends on one or two customers is exposed to the loss of either. (Moderate) — Supplies the general principle for exposure but nothing on likelihood, magnitude or mitigants. It does not support the comparative 'main vulnerability' claim or the cash-fragility claim.

View this argument on LogicFirst.ai